LOCAL PREVIEW View on GitHub

Continual Learning Scenarios - MangaAssist

Continual learning keeps MangaAssist current as manga trends, catalog inventory, policies, and user language change. The challenge is improving on new data without forgetting old behavior that still matters.

When This Topic Matters

Use continual learning when:

  • new manga releases change search and recommendation patterns,
  • seasonal promotions alter traffic mix,
  • users adopt new slang,
  • policy updates change support answers,
  • a retrained model regresses on old intents.

Scenario 1 - Monthly Intent Refresh

The intent classifier already handles 10 known intents. Each month, collect:

  • low-confidence requests,
  • corrected routing labels,
  • support outcome labels,
  • rejected OOD clusters,
  • sampled high-confidence predictions for regression checks.

Training strategy:

Component Purpose
new labeled data learn current behavior
rehearsal buffer preserve older intent patterns
golden set block regressions on critical workflows
drift report decide whether retraining is needed

Promotion gate:

Metric Gate
new-month accuracy improves by >= 1 point
old golden-set accuracy no drop over 0.3 points
rare-class recall no critical regression
business-weighted harm improves or stays within budget

Scenario 2 - New Release Drift

Example event:

A major anime adaptation launches. Users suddenly search for a newly popular title,
its spin-offs, and related merch.

Risk:

  • product discovery traffic spikes,
  • embedding retrieval overweights old best sellers,
  • recommendation model misses new comparisons.

Continual update:

  • add recent clicks and purchases,
  • add editorial pairs for the new title,
  • refresh retrieval adapter monthly,
  • keep an old-title validation set to avoid popularity collapse.

Scenario 3 - Policy Change Without Forgetting

If return policy changes, support answers need current behavior, but the model must still handle older orders correctly.

Do not solve this only with fine-tuning. Use retrieval-grounded policy context first. Continual fine-tuning can teach answer structure, but the exact policy should come from current retrieved documents.

Failure Modes

Failure Detection Fix
catastrophic forgetting old golden-set drop rehearsal buffer and lower LR
trend overfitting model over-routes to new title balanced sampling
stale support behavior old policy appears in answer RAFT with current policy docs
unnecessary retraining cost rises without quality gain require drift plus accuracy evidence

Production Log

{
  "event": "continual_learning_decision",
  "window": "2026-04",
  "kl_divergence": 0.031,
  "sampled_accuracy": 0.898,
  "golden_set_accuracy": 0.921,
  "decision": "retrain_with_rehearsal"
}

Final Decision

For MangaAssist, continual learning should be evidence-driven. Retrain when drift creates measurable harm, then protect old behavior with rehearsal data, golden sets, and business-weighted gates.